A method and system for constructing and analyzing multi-omics mass spectrometry profiles of thyroid diseases

Abstract

The present application provides a method and system for constructing and analyzing multi-omics mass spectrometry maps of thyroid diseases, wherein thyroid ultrasound images and multi-omics mass spectrometry data are obtained, and cross-modal correlation relationships are generated based on the anatomical structural features of the thyroid ultrasound images and the spatial distribution features of the biomolecules of the multi-omics mass spectrometry data. In combination with the thyroid anatomical deformation field and multi-task optimization objectives, nonlinear coupling calculations are performed on the back-propagation error during the multi-task learning iteration process to obtain the gradient contribution weights between tasks; based on the cross-modal correlation features, the anatomical deformation field, and the gradient contribution weights, a target multi-omics mass spectrometry map is generated, and the thyroid pathological characteristics are analyzed to obtain the three-dimensional positioning coordinates of the thyroid nodules and the predicted value of the malignancy probability. The present application improves the comprehensiveness and accuracy of thyroid disease diagnosis.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical fields of medical image analysis, multi-omics data fusion, and artificial intelligence-assisted diagnosis, and in particular, to a method and system for constructing and analyzing multi-omics mass spectrometry maps of thyroid diseases. Background Art

[0002] Accurately identifying the location and malignancy of thyroid nodules is crucial for the diagnosis and treatment of thyroid diseases. Currently, the medical field urgently needs a technology that can integrate multimodal data (such as ultrasound images and mass spectrometry data) to improve diagnostic accuracy and efficiency. This technology needs to be able to process and analyze data from various sources and generate reliable pathological analysis results to assist physicians in making more accurate diagnoses and treatment decisions.

[0003] To address this technological need, a multimodal data fusion method based on deep learning has been proposed. This method constructs a deep learning model to extract and fuse features from thyroid ultrasound images and mass spectrometry data, thereby enabling the localization of thyroid nodules and the prediction of malignancy. Specifically, this method utilizes a convolutional neural network (CNN) to extract features from ultrasound images, while simultaneously processing the mass spectrometry data using a specific algorithm. Finally, these features are integrated through a multi-task learning framework to achieve nodule localization and malignancy prediction.

[0004] While existing approaches have made progress in localizing and predicting thyroid nodules, they still face limitations when processing multimodal data. Specifically, when fusing ultrasound images and mass spectrometry data, existing approaches fail to fully consider the cross-modal correlation between the two. This results in significant discrepancies between the generated mass spectrometry profiles and the anatomical features in the ultrasound images. This discrepancy not only affects the accuracy of nodule localization but also reduces the reliability of malignancy prediction, thereby limiting the effectiveness of these approaches in clinical practice. Summary of the Invention

[0005] The embodiments of the present application provide a method and system for constructing and analyzing a multi-omics mass spectrometry map of thyroid disease, which is used to solve the problems of poor comprehensiveness and low accuracy in the diagnosis of thyroid disease in the prior art.

[0006] In a first aspect, the present invention provides a method for constructing and analyzing a multi-omics mass spectrometry profile of a thyroid disease, comprising:

[0007] Acquiring a thyroid ultrasound image and multi-omics mass spectrometry data, and generating a cross-modal association relationship based on anatomical structural features of the thyroid ultrasound image and biomolecule spatial distribution features of the multi-omics mass spectrometry data, wherein the anatomical structure includes a thyroid nodule;

[0008] generating a thyroid anatomical structure deformation field based on the anatomical structure features, and constructing a multi-task optimization target according to preset thyroid disease diagnosis requirements, wherein the multi-task optimization target includes a thyroid nodule localization task and a thyroid nodule malignancy prediction task;

[0009] During the multi-task learning iteration process, a nonlinear coupling calculation is performed on the back propagation error of the thyroid anatomical structure deformation field to obtain the respective gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task;

[0010] generating a cross-modal association feature according to the cross-modal association relationship, and generating a target multi-omics mass spectrum based on the thyroid anatomical structure deformation field and the cross-modal association feature in combination with the gradient contribution weight, wherein the error between the metabolomics signal hotspot area in the target multi-omics mass spectrum and the spatial coordinates of the microcalcification cluster in the thyroid ultrasound image is within a preset anatomical structure tolerance range;

[0011] The target multi-omics mass spectrometry map is analyzed for thyroid pathological characteristics to obtain thyroid pathological analysis results, which include the three-dimensional positioning coordinates of the thyroid nodule and the predicted value of the malignancy probability.

[0012] Optionally, generating a cross-modal association feature according to the cross-modal association relationship, and generating a target multi-omics mass spectrum based on the thyroid anatomical structure deformation field and the cross-modal association feature in combination with the gradient contribution weight, includes:

[0013] Based on the cross-modal association relationship, the anatomical structure features and the biomolecule spatial distribution features are combined to construct a three-dimensional spatial weight matrix, and the anatomical structure features are subjected to multi-level spatial decomposition according to the three-dimensional spatial weight matrix to obtain texture distribution patterns at different anatomical levels in the thyroid tissue;

[0014] fusing the texture distribution pattern with the multi-omics mass spectrometry data layer by layer to generate cross-modal correlation features;

[0015] performing task-aware filtering on the cross-modal correlation feature according to the gradient contribution weight to obtain a filtered cross-modal correlation feature;

[0016] Nonlinearly superimposing the filtered cross-modal correlation features with the thyroid anatomical structure deformation field to generate an initial multi-omics mass spectrum;

[0017] The initial multi-omics mass spectrum is optimized to generate a target multi-omics mass spectrum.

[0018] Optionally, fusing the texture distribution pattern with the multi-omics mass spectrometry data layer by layer to generate cross-modal correlation features includes:

[0019] Analyzing the multi-omics mass spectrometry data to obtain mass spectrometry detection points, and dynamically adjusting the mass spectrometry detection points within a coordinate mapping range of each level based on the hierarchical resolution of the thyroid ultrasound image, wherein the global coordinate mapping range covers the overall contour of the thyroid gland and the topological boundaries of adjacent organs, and the local level mapping range is limited to the local spatial neighborhood of the microcalcification cluster in the thyroid ultrasound image;

[0020] At each level of resolution, performing spatial interpolation compensation processing on the metabolomics signal intensity distribution in the multi-omics mass spectrometry data according to the adjusted spatial distance between the mass spectrometry detection point and the ultrasound image voxel to obtain a compensated metabolomics signal intensity distribution, wherein the ultrasound image voxel is the smallest spatial unit of the thyroid ultrasound image;

[0021] performing directionality enhancement processing on the texture features of the texture distribution pattern according to the spatial consistency between the spatial distribution path of the proteomic marker expression amount in the multi-omics mass spectrometry data and the directional features of the texture distribution pattern to obtain enhanced texture features;

[0022] The compensated metabolomics signal intensity distribution and the enhanced texture features are channel-spliced according to the hierarchical resolution to generate cross-modal correlation features, wherein the global-level splicing channel contains the covariance matrix of the overall thyroid deformation trend and the global distribution of the metabolomics signal, and the local-level splicing channel contains the texture details of the microcalcification cluster and the spatial offset of the peak position of the metabolic signal.

[0023] Optionally, the nonlinearly superimposing the filtered cross-modal correlation features with the thyroid anatomical structure deformation field to generate an initial multi-omics mass spectrum includes:

[0024] Decomposing the deformation field of the thyroid anatomical structure into local deformation parameters and global deformation trends within the thyroid tissue;

[0025] Based on the local deformation parameters, performing spatial coordinate transformation on the filtered cross-modal correlation features to obtain cross-modal correlation features after spatial coordinate transformation;

[0026] Based on the global deformation trend, performing inter-level weight distribution on the cross-modal correlation features after the spatial coordinate transformation to obtain cross-modal correlation features after weight distribution;

[0027] The weighted cross-modal correlation features are nonlinearly superimposed with the deformation parameters of the thyroid anatomical structure deformation field, and the spatial coordinates of the metabolomics signal hotspot area and the microcalcification cluster are iteratively optimized to generate an initial multi-omics mass spectrum map.

[0028] Optionally, the target multi-omics mass spectrometry is subjected to thyroid pathology feature analysis to obtain thyroid pathology analysis results, wherein the thyroid pathology analysis results include the three-dimensional positioning coordinates of the thyroid nodules and the predicted value of the malignancy probability, including:

[0029] Decomposing the thyroid anatomical structure in the target multi-omics mass spectrometry map into global hierarchical anatomical deformation features and local hierarchical microcalcification cluster density features. During the decomposition process, the global hierarchical anatomical deformation features are generated by quantifying the displacement gradient direction of the interface between the thyroid margin and the carotid artery, and the local hierarchical microcalcification cluster density features are obtained by statistically analyzing the three-dimensional coordinate distribution density of microcalcification clusters within a preset spatial neighborhood.

[0030] Extracting a candidate thyroid nodule region from an overlapping region between the metabolomics signal hotspot region and the local hierarchical microcalcification cluster density feature, and generating an initial boundary contour based on the candidate thyroid nodule region and a consistency coefficient between the metabolomics signal gradient direction and the texture distribution pattern direction;

[0031] Based on the expression levels of the proteomic markers and in combination with the global hierarchical anatomical deformation characteristics, a set of mutation positions associated with a preset malignancy risk is screened from the candidate thyroid nodule region, and based on the set of mutation positions, a hotspot of abnormal protein expression is generated;

[0032] Calculating the malignancy risk direction confidence of the thyroid nodule candidate region according to the spatial angle between the metabolomics signal gradient direction and the texture distribution pattern;

[0033] The spatial distribution path of the abnormal protein expression hotspot is superimposed with the confidence of the malignant risk direction to generate the three-dimensional positioning coordinates of the thyroid nodule and the malignancy probability prediction value.

[0034] Optionally, superimposing the spatial distribution path of the abnormal protein expression hotspot and the malignancy risk direction confidence to generate the three-dimensional positioning coordinates of the thyroid nodule and the malignancy probability prediction value includes:

[0035] Generate a spatial distribution path of the abnormal protein expression hotspot by calculating the similarity between the spatial distances between the mutation positions in the mutation position set and the expression amount of the proteomic marker;

[0036] Superimposing the spatial distribution path and the malignant risk direction confidence to generate a superposition result;

[0037] Based on the superposition result, the coupling error between the spatial distribution path and the initial boundary contour is optimized to determine the three-dimensional positioning coordinates of the thyroid nodule, and the spatial correlation between the spatial distribution path and the texture distribution pattern is verified to generate a predicted value of the malignancy probability of the thyroid nodule.

[0038] Optionally, in the multi-task learning iteration process, nonlinear coupling calculation is performed on the back propagation error of the thyroid anatomical structure deformation field to obtain the respective gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task, including:

[0039] Decomposing the back propagation error of the thyroid anatomical structure deformation field into a back propagation error component for the thyroid nodule localization task and a back propagation error component for the thyroid nodule malignancy prediction task;

[0040] constructing a nonlinear coupling function based on a nonlinear relationship between a back-propagation error component of the thyroid nodule localization task and a back-propagation error component of the thyroid nodule malignancy prediction task;

[0041] Based on the nonlinear coupling function, the gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task are calculated.

[0042] In a second aspect, the present invention provides a multi-omics mass spectrometry analysis system for thyroid disease, comprising:

[0043] an acquisition module, configured to acquire thyroid ultrasound images and multi-omics mass spectrometry data, and generate a cross-modal association relationship based on the anatomical structural features of the thyroid ultrasound images and the spatial distribution features of biomolecules in the multi-omics mass spectrometry data;

[0044] A construction module is used to generate a thyroid anatomical structure deformation field based on the anatomical structure features, and to construct a multi-task optimization target according to preset thyroid disease diagnosis requirements, wherein the multi-task optimization target includes a thyroid nodule localization task and a thyroid nodule malignancy prediction task;

[0045] a calculation module, configured to perform nonlinear coupling calculation on the back propagation error of the thyroid anatomical structure deformation field during the multi-task learning iteration process, to obtain the gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task;

[0046] a generation module, configured to generate a cross-modal association feature according to the cross-modal association relationship, and generate a target multi-omics mass spectrum based on the thyroid anatomical structure deformation field and the cross-modal association feature, in combination with the gradient contribution weight, wherein the error between the metabolomics signal hotspot area in the target multi-omics mass spectrum and the spatial coordinates of the microcalcification cluster in the thyroid ultrasound image is within a preset anatomical structure tolerance range;

[0047] An analysis module is used to analyze the thyroid pathological characteristics of the target multi-omics mass spectrometry map to obtain thyroid pathological analysis results, which include the three-dimensional positioning coordinates of the thyroid nodules and the predicted value of the malignancy probability.

[0048] In a third aspect, an embodiment of the present application provides a computing device comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute a multi-omics mass spectrometry analysis method for thyroid disease as described in any one of the first aspects.

[0049] In a fourth aspect, an embodiment of the present application provides a computer storage medium having computer program instructions stored thereon, which, when executed by a processor, implements a method for constructing and analyzing a multi-omics mass spectrometry map of thyroid disease as described in any one of the first aspects.

[0050] In an embodiment of the present application, a method for constructing and analyzing a multi-omics mass spectrometry map of a thyroid disease is provided, the method comprising: acquiring a thyroid ultrasound image and multi-omics mass spectrometry data, generating a cross-modal correlation relationship based on the anatomical structure characteristics of the thyroid ultrasound image and the biomolecular spatial distribution characteristics of the multi-omics mass spectrometry data, wherein the anatomical structure includes a thyroid nodule; generating a thyroid anatomical structure deformation field based on the anatomical structure characteristics, and constructing a multi-task optimization target according to preset thyroid disease diagnosis requirements, wherein the multi-task optimization target includes a thyroid nodule positioning task and a thyroid nodule malignancy prediction task; in the multi-task learning iteration process, performing nonlinear coupling on the back propagation error of the thyroid anatomical structure deformation field. Combined calculation is performed to obtain the gradient contribution weights of the thyroid nodule positioning task and the thyroid nodule malignancy prediction task respectively; cross-modal association features are generated according to the cross-modal association relationship, and based on the thyroid anatomical structure deformation field and the cross-modal association features, combined with the gradient contribution weights, a target multi-omics mass spectrum is generated, and the error between the metabolomics signal hotspot area in the target multi-omics mass spectrum and the spatial coordinates of the microcalcification cluster in the thyroid ultrasound image is within the preset anatomical structure tolerance range; thyroid pathological characteristics are analyzed on the target multi-omics mass spectrum to obtain thyroid pathological analysis results, which include the three-dimensional positioning coordinates of the thyroid nodule and the malignancy probability prediction value.

[0051] The technical solution of this application has the following beneficial effects:

[0052] This application realizes the acquisition of multimodal data, providing a basis for subsequent cross-modal data fusion. The association between the anatomical structural features of ultrasound images and the biomolecular features of multi-omics mass spectrometry data is established, solving the problem of isolated multimodal data. The deformation characteristics of thyroid tissue are quantified to provide spatial deformation information for multi-task optimization. The optimization goals of thyroid nodule localization and malignancy prediction tasks are clarified to support the collaborative optimization of multi-task learning. Through nonlinear coupling calculation, the gradient contribution weights between tasks are dynamically adjusted to solve the gradient conflict problem between tasks. Efficient fusion and spatial alignment of multimodal data are achieved to generate high-precision multi-omics mass spectrometry maps. Based on the multi-omics mass spectrometry map, the three-dimensional positioning coordinates and malignancy probability prediction values of thyroid nodules are generated, which improves the comprehensiveness and accuracy of the diagnosis.

[0053] Furthermore, the embodiments of the present application also construct a three-dimensional spatial weight matrix based on the cross-modal correlation relationship, combining anatomical structure characteristics and biological molecular spatial distribution characteristics, and perform multi-level spatial decomposition of anatomical structure characteristics to obtain texture distribution patterns at different anatomical levels in thyroid tissue; the texture distribution patterns are fused with multi-omics mass spectrometry data layer by layer to generate cross-modal correlation features; the cross-modal correlation features are task-aware filtered according to the gradient contribution weight to obtain filtered cross-modal correlation features; the filtered cross-modal correlation features are nonlinearly superimposed with the thyroid anatomical structure deformation field to generate an initial multi-omics mass spectrometry map; the initial multi-omics mass spectrometry map is optimized to generate a target multi-omics mass spectrometry map.

[0054] By constructing a three-dimensional spatial weight matrix and multi-level spatial decomposition, efficient fusion of thyroid anatomical structure characteristics and multi-omics mass spectrometry data is achieved; through task-aware filtering and nonlinear superposition, the spatial alignment accuracy of cross-modal correlation features and anatomical structure deformation fields is optimized; the target multi-omics mass spectrometry map finally generated improves the comprehensiveness and accuracy of thyroid disease diagnosis and solves the problems of isolated multimodal data and insufficient spatial alignment accuracy.

[0055] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0057] Figure 1A flowchart of a method for constructing and analyzing a multi-omics mass spectrometry profile of thyroid disease provided in an embodiment of the present application;

[0058] Figure 2 A schematic diagram of the structure of a multi-omics mass spectrometry analysis system for thyroid disease provided in an embodiment of the present application;

[0059] Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to enable people skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0061] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0062] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0063] This application acquires thyroid ultrasound images and multi-omics mass spectrometry data, and generates cross-modal correlation relationships based on the anatomical structural features of thyroid ultrasound images (including thyroid nodules) and the spatial distribution features of biomolecules in multi-omics mass spectrometry data; generates a thyroid anatomical deformation field based on the anatomical structural features, and constructs a multi-task optimization target (including thyroid nodule positioning task and malignancy prediction task) according to the needs of thyroid disease diagnosis; in the multi-task learning iteration process, performs nonlinear coupling calculation on the back propagation error of the thyroid anatomical deformation field to obtain the gradient contribution weight of each task; generates cross-modal correlation features based on the cross-modal correlation relationship, and combines the thyroid anatomical deformation field and the gradient contribution weight to generate a target multi-omics mass spectrometry map to ensure that the spatial coordinate error between the metabolomics signal hotspot area and the microcalcification cluster in the thyroid ultrasound image is within the preset anatomical structure tolerance range; performs thyroid pathological feature analysis on the target multi-omics mass spectrometry map to obtain thyroid pathological analysis results (including the three-dimensional positioning coordinates of thyroid nodules and the malignancy probability prediction value), thereby achieving accurate diagnosis and prediction of thyroid diseases.

[0064] Figure 1 A flowchart of a method for constructing and analyzing a multi-omics mass spectrometry map of thyroid disease is provided in the present application embodiment. Figure 1 As shown, the method includes:

[0065] Step 101: Acquire a thyroid ultrasound image and multi-omics mass spectrometry data, and generate a cross-modal correlation relationship based on the anatomical structure characteristics of the thyroid ultrasound image and the biomolecule spatial distribution characteristics of the multi-omics mass spectrometry data.

[0066] In this step, the multi-omics mass spectrometry data includes the distribution of metabolomics signal intensity and the expression of proteomics markers. Cross-modal associations are used to fuse data from different modalities. Anatomical structures include thyroid nodules. Thyroid ultrasound images are images of the thyroid structure obtained using ultrasound imaging technology, which are used to display anatomical structural features such as the morphology, size, location, and nodules of the thyroid gland. For example, the image can clearly show the boundaries, internal echoes, and blood flow of thyroid nodules. The spatial distribution characteristics of biomolecules refer to the spatial distribution patterns of biomolecules such as metabolomics and proteomics in multi-omics mass spectrometry data. For example, metabolomics signals are highly concentrated in the thyroid nodule area.

[0067] In practice, thyroid ultrasound images and multi-omics mass spectrometry data are acquired, and a cross-modal correlation is generated based on the anatomical structural features of the thyroid ultrasound image and the spatial distribution features of biomolecules in the multi-omics mass spectrometry data. Specifically, feature alignment technology is used to correlate the anatomical structural features of the ultrasound image with the biomolecule distribution features of the mass spectrometry data to generate a cross-modal correlation.

[0068] For example, a thyroid ultrasound image of patient A showed a 1.5 cm × 1.0 cm nodule located in the right lobe. Multi-omics mass spectrometry data revealed a high concentration of metabolomics signals in the right lobe. Spatial registration technology was used to align the ultrasound image with the mass spectrometry data, generating cross-modal correlations.

[0069] Step 102: Generate a thyroid anatomical structure deformation field based on the anatomical structure features, and construct a multi-task optimization target according to preset thyroid disease diagnosis requirements.

[0070] In this step, the thyroid anatomical deformation field is used to describe the deformation and spatial distribution of the thyroid structure. The multi-task optimization objectives include the thyroid nodule localization task and the thyroid nodule malignancy prediction task, which are used to guide multi-task learning. Thyroid disease diagnosis requirements include the precise localization and malignancy prediction of thyroid nodules. Specifically, the precise localization requirement requires accurate identification of the three-dimensional coordinates of the nodule, and the malignancy prediction requirement requires an assessment of the malignancy probability of the nodule. For example, diagnostic requirements may include a nodule localization error of <0.1 cm and a malignancy prediction accuracy of >90%.

[0071] In practice, a thyroid anatomical deformation field is generated based on these anatomical features, and a multi-task optimization objective is constructed based on the pre-defined thyroid disease diagnosis requirements. Specifically, a deformation model (such as elastic registration) is used to generate the thyroid anatomical deformation field, and a multi-task optimization objective (such as mean squared error for nodule localization and cross-entropy loss for malignancy prediction) is constructed.

[0072] For example, based on the thyroid ultrasound image of patient A, an anatomical structure deformation field is generated to describe the deformation characteristics of the nodule area; a multi-task optimization goal is constructed, including the nodule localization task (target error <0.1cm) and the malignancy prediction task (target accuracy >90%).

[0073] Step 103: During the multi-task learning iteration process, a nonlinear coupling calculation is performed on the back propagation error of the thyroid anatomical structure deformation field to obtain the respective gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task.

[0074] In this step, the gradient contribution weight refers to the proportion of contributions to model training from the thyroid nodule localization task and the malignancy prediction task in multi-task learning. This weight is used to dynamically adjust the optimization objective weights between the tasks. In practice, during the iterative multi-task learning process, a nonlinear coupling calculation is performed on the back-propagated error of the thyroid anatomical structure deformation field to obtain the gradient contribution weights for each of the thyroid nodule localization and malignancy prediction tasks. Specifically, the gradient contribution weights for each task are calculated using a nonlinear coupling algorithm (e.g., weighted summation).

[0075] For example, in a multi-task learning iteration, the gradient contribution weight for the nodule localization task is calculated as 0.6, and the gradient contribution weight for the malignancy prediction task is calculated as 0.4.

[0076] Step 104: Generate a cross-modal association feature based on the cross-modal association relationship, and generate a target multi-omics mass spectrum based on the thyroid anatomical structure deformation field and the cross-modal association feature, combined with the gradient contribution weight, wherein the error between the metabolomics signal hotspot area in the target multi-omics mass spectrum and the spatial coordinates of the microcalcification cluster in the thyroid ultrasound image is within a preset anatomical structure tolerance range.

[0077] In this step, the target multi-omics mass spectrum contains the concentration distribution of metabolomics-protein bi-omics biomarkers, the three-dimensional positioning coordinates of the nodules, and the probability value of malignant risk, which are used to accurately locate the hot spots of metabolomics signals. The spatial coordinate error is an explicit indicator of cross-modal data alignment, while the back propagation error is an implicit indicator of model training optimization. Cross-modal correlation features are used to describe the correlation between the two modal data. For example, the correlation between the boundary features of the nodules in the ultrasound image and the spatial distribution characteristics of the metabolomics signals in the mass spectrometry data. The metabolomics signal hotspot refers to the area with high metabolomics signal intensity in thyroid tissue, which reflects the spatial distribution pattern of metabolite concentration.

[0078] In practice, a cross-modal correlation feature is generated based on the cross-modal correlation relationship. A target multi-omics mass spectrum is generated based on the thyroid anatomical deformation field and the cross-modal correlation feature, combined with the gradient contribution weights. Specifically, feature fusion techniques (such as weighted fusion) are used to generate the target multi-omics mass spectrum, ensuring that the spatial coordinate error between the metabolomics signal hotspots and the microcalcification clusters in the thyroid ultrasound image is within a preset anatomical structure tolerance.

[0079] For example, based on the cross-modal correlation features and anatomical structure deformation field of patient A, a targeted multi-omics mass spectrometry map was generated, in which the spatial coordinate error between the metabolomics signal hotspot area and the microcalcification cluster in the ultrasound image was <0.1 cm.

[0080] Step 105: Analyze the thyroid pathology characteristics of the target multi-omics mass spectrum to obtain thyroid pathology analysis results.

[0081] In this step, the thyroid pathology analysis results include the 3D coordinates of the thyroid nodule and a predicted malignancy probability. For example, the analysis results may include the 3D coordinates of the nodule (1.2 cm, 0.8 cm, 0.5 cm) and a predicted malignancy probability of 85%. The 3D coordinates are spatially corrected by fusing the gradient change direction of the metabolomics signal intensity distribution with the texture expansion pattern of the hypoechoic region in the thyroid ultrasound image. The predicted malignancy probability is generated by coupling the location of the proteomic marker expression mutation with the anatomical deformation characteristics at the thyroid-carotid artery interface to generate a risk confidence index.

[0082] In practice, the target multi-omics mass spectrum is analyzed for thyroid pathological features to obtain thyroid pathological analysis results. Specifically, the target multi-omics mass spectrum is analyzed using pathological feature extraction technology (such as a deep learning model) to generate the three-dimensional location coordinates of the thyroid nodule and a predicted value for the probability of malignancy.

[0083] For example, the pathological characteristics of the target multi-omics mass spectrometry map of patient A were analyzed, and the three-dimensional positioning coordinates of the thyroid nodule were obtained as (1.2cm, 0.8cm, 0.5cm), and the predicted value of the malignancy probability was 85%.

[0084] By acquiring thyroid ultrasound images and multi-omics mass spectrometry data, cross-modal correlation relationships and thyroid anatomical structure deformation fields are generated, and multi-task optimization targets are constructed; during the multi-task learning iteration process, gradient contribution weights are generated through nonlinear coupling calculations, and target multi-omics mass spectrometry maps are generated in combination with cross-modal correlation features; finally, thyroid pathology analysis results are generated through pathological feature analysis, achieving accurate diagnosis and prediction of thyroid diseases and improving the accuracy and reliability of diagnosis.

[0085] To address the issues of isolated multimodal data and insufficient spatial alignment accuracy in thyroid disease diagnosis, this application constructs a three-dimensional spatial weight matrix based on cross-modal association relationships, combining the anatomical structural features of thyroid ultrasound images with the spatial distribution characteristics of biomolecules in multi-omics mass spectrometry data. This matrix quantifies the feature matching between ultrasound image voxels and mass spectrometry detection points. Through multi-level spatial decomposition, texture distribution patterns at different anatomical levels within thyroid tissue are extracted (including the overall contour at the global level, the regional nodule distribution at the mesoscopic level, and the microcalcification cluster texture at the local level). The texture distribution patterns are fused layer by layer with the multi-omics mass spectrometry data to generate cross-modal association features. Task-aware filtering is performed on the cross-modal association features based on gradient contribution weights to dynamically adjust the feature contribution ratios for the thyroid nodule localization task and the malignancy prediction task. The filtered cross-modal association features are nonlinearly superimposed with the thyroid anatomical deformation field to generate an initial multi-omics mass spectrometry map. The target multi-omics mass spectrometry map is generated by iteratively optimizing the spatial coordinate errors between metabolomics signal hotspots and microcalcification clusters, achieving efficient fusion and spatial alignment of multimodal data.

[0086] In some embodiments, generating a cross-modal correlation feature according to the cross-modal correlation relationship in step 104, and generating a target multi-omics mass spectrum based on the thyroid anatomical structure deformation field and the cross-modal correlation feature in combination with the gradient contribution weight, includes:

[0087] Step 201: Based on the cross-modal association relationship, combined with the anatomical structure characteristics and the biological molecule spatial distribution characteristics, a three-dimensional spatial weight matrix is constructed, and the anatomical structure characteristics are subjected to multi-level spatial decomposition according to the three-dimensional spatial weight matrix to obtain texture distribution patterns at different anatomical levels in the thyroid tissue.

[0088] In step 201, the three-dimensional spatial weight matrix is a three-dimensional tensor, each element of which represents the degree of feature matching between an ultrasound image voxel and a mass spectrometry detection point at a specific spatial location (x, y, z). This matrix is used to quantify the spatial consistency of multimodal data. Multi-level spatial decomposition involves performing multi-resolution analysis of the anatomical features of the thyroid ultrasound image to extract texture distribution patterns at different levels. Decomposition process: The three-dimensional spatial weight matrix is applied to the anatomical features of the ultrasound image to perform multi-level spatial decomposition. For example, the weight matrix and the anatomical features of the ultrasound image are decomposed layer by layer through a convolution operation to extract texture distribution patterns at different levels. Based on the distribution of the weight matrix, texture distribution patterns at different anatomical levels within the thyroid tissue are extracted. For example, the overall contour at the global level, the regional nodule distribution at the mesoscopic level, and the microcalcification cluster texture at the local level are extracted. Specific numerical example: Assuming that the physical spatial range of the thyroid region is 0-100 mm on the x-axis, 0-100 mm on the y-axis, and 0-50 mm on the z-axis, a 100×100×50 three-dimensional spatial weight matrix is constructed, where each element A represents the feature matching degree between the ultrasound image voxel and the mass spectrometry detection point at the spatial position (i, j, k). For example, the weight value of the position (50, 60, 20) is 0.12, the weight value of the position (55, 65, 22) is 0.15, and the weight value of the position (70, 80, 25) is 0.08. Based on this weight matrix, the anatomical structural features of the thyroid ultrasound image are decomposed into multiple levels of space: at the global level, the weight matrix is low-pass filtered and multiplied element-by-element by the low-resolution features of the ultrasound image (64×64×16). For example, the weight value of position (25, 30, 8) is 0.10, and the texture intensity is 0.8. The value of this position in the global texture distribution pattern is 0.10×0.8=0.080.10×0.8=0.08, reflecting the topological relationship between the overall contour of the thyroid gland and its neighboring organs. At the meso-level, the weight matrix is band-pass filtered and multiplied element-by-element by the medium-resolution features of the ultrasound image (128×128×32). For example, the weight value of position (50, 60, 16) is 0.10, and the texture intensity is 0.8. The weight value is 0.12, and the texture intensity is 0.6. The value of this position in the mesoscopic texture distribution pattern is 0.12×0.6=0.0720.12×0.6=0.072, reflecting the regional nodule distribution and blood flow signal. At the local level, the weight matrix is high-pass filtered and multiplied element-by-element by the high-resolution features of the ultrasound image (256×256×64). For example, the weight value of position (100, 120, 32) is 0.15, and the texture intensity is 0.9. The value of this position in the local texture distribution pattern is 0.15×0.9=0.1350.15×0.9=0.135, reflecting the texture details of the microcalcification cluster.The final output is texture distribution patterns at the global, meso, and local levels, describing the overall thyroid contour, regional nodule distribution, and microcalcification cluster texture, respectively. Texture distribution patterns refer to texture features at different anatomical levels within thyroid ultrasound images, including the overall contour at the global level, regional nodule distribution at the meso level, and microcalcification cluster texture at the local level. Thyroid tissue refers to the physiological structure of the thyroid gland, including its cells, blood vessels, and nerves. For example, thyroid tissue is composed of thyroid follicular cells and interstitial tissue, responsible for secreting thyroid hormones. Thyroid anatomy refers to the overall and local anatomical features of the thyroid gland, including its margins, internal structures, and topological relationships with adjacent organs (such as the carotid artery and trachea). Examples include the overall contour of the thyroid gland, regional blood flow distribution, and local microcalcification cluster texture. Thyroid nodules are areas of localized abnormal proliferation within the thyroid tissue, typically appearing as hypoechoic or hyperechoic areas on ultrasound images, and can be benign or malignant. For example, thyroid nodules can appear as masses or cysts within the thyroid tissue. The thyroid anatomy is the overall and local anatomical features of the thyroid gland, while thyroid nodules are local pathological features within the thyroid anatomy. Thyroid tissue is the physiological basis of thyroid anatomy, and thyroid anatomy describes the spatial and functional aspects of thyroid tissue. Thyroid nodules are local pathological features within thyroid tissue and are also part of the thyroid anatomy.

[0089] In an embodiment of the present application, first, based on the cross-modal association relationship, a three-dimensional spatial weight matrix is constructed by combining the anatomical structural characteristics of the thyroid ultrasound image and the spatial distribution characteristics of the biomolecules of the multi-omics mass spectrometry data. Specifically, each element of the weight matrix is generated by calculating the spatial distance and feature similarity between the ultrasound image voxels and the mass spectrometry detection points. Then, the anatomical structural characteristics of the thyroid ultrasound image are subjected to multi-level spatial decomposition according to the three-dimensional spatial weight matrix to extract the overall outline of the global level, the regional nodule distribution at the mesoscopic level, and the microcalcification cluster texture at the local level. Finally, the texture distribution pattern of different anatomical levels in the thyroid tissue is obtained.

[0090] Step 202: Fusing the texture distribution pattern with the multi-omics mass spectrometry data layer by layer to generate cross-modal correlation features.

[0091] In step 202 , layer-by-layer fusion refers to aligning and fusing the texture distribution pattern with the multi-omics mass spectrometry data at a hierarchical resolution to generate cross-modal correlation features.

[0092] In this application example, texture distribution patterns are fused layer by layer with multi-omics mass spectrometry data. Specifically, at the global level, the overall thyroid contour is aligned with the global distribution of metabolomics signals; at the meso-level, the regional nodule distribution is aligned with the regional distribution of metabolomics signals; and at the local level, the texture of microcalcification clusters is aligned with the local peaks of metabolomics signals. Through layer-by-layer fusion, cross-modal correlation features are generated, achieving efficient fusion of multimodal data.

[0093] Step 203: Perform task-aware filtering on the cross-modal correlation features according to the gradient contribution weights to obtain filtered cross-modal correlation features.

[0094] In step 203, the task-aware filtering dynamically adjusts the contribution ratio of the feature channels of the cross-modal correlation feature tensor between the nodule localization task and the malignancy prediction task, wherein the feature channel of the nodule localization task retains the spatial coupling information of the texture extension pattern inside the thyroid tissue and the direction of the metabolomics signal gradient change, and the feature channel of the malignancy prediction task retains the temporal correlation information of the proteomic marker expression mutation position and the deformation characteristics of the thyroid-carotid artery interface.

[0095] In the embodiments of the present application, task-aware filtering is performed on cross-modal correlation features based on gradient contribution weights. Specifically, the gradient contribution weights between the tasks are dynamically adjusted by calculating the back-propagation error between the thyroid nodule localization task and the malignancy prediction task. The cross-modal correlation features are then filtered based on the gradient contribution weights, retaining the features that contribute most to the thyroid nodule localization and malignancy prediction tasks, and generating filtered cross-modal correlation features.

[0096] Step 204: nonlinearly superimpose the filtered cross-modal correlation features with the thyroid anatomical structure deformation field to generate an initial multi-omics mass spectrum.

[0097] In step 204 , nonlinear superposition refers to weighted fusion of the filtered cross-modal correlation features and the thyroid anatomical structure deformation field.

[0098] In this example, filtered cross-modal correlation features were nonlinearly superimposed with the thyroid anatomical deformation field. Specifically, an initial multi-omics mass spectrometry map was generated by weighted fusion of the cross-modal correlation features and deformation parameters. The spatial alignment accuracy of the initial map was then corrected by iteratively optimizing the spatial coordinate errors between metabolomics signal hotspots and microcalcification clusters.

[0099] Step 205: Optimize the initial multi-omics mass spectrum to generate a target multi-omics mass spectrum.

[0100] In step 205, the optimization is performed by constraining the spatial correlation threshold between the mutation position of the proteomic marker expression and the texture extension pattern of the low-echo area in the ultrasound image, while limiting the deformation amplitude of the nodule three-dimensional positioning coordinates and the thyroid anatomical structure deformation field within a preset tolerance range.

[0101] In this embodiment of the present application, the initial multi-omics mass spectrometry profile was optimized. Specifically, a target multi-omics mass spectrometry profile was generated by constraining the spatial correlation threshold between the location of proteomic marker expression mutations and the texture expansion pattern of the hypoechoic region, and limiting the deformation amplitude of the nodule's three-dimensional positioning coordinates and the thyroid anatomical structure deformation field to within a preset tolerance range.

[0102] Here's a specific example:

[0103] Assuming the physical spatial extent of the thyroid region is 0–100 mm along the x-axis, 0–100 mm along the y-axis, and 0–50 mm along the z-axis, a 100×100×50 three-dimensional spatial weight matrix was constructed. At the global level, the overall thyroid contour was aligned with the global distribution of the metabolomics signal to generate a global texture distribution pattern. At the meso-level, the regional nodule distribution was aligned with the regional distribution of the metabolomics signal to generate a meso-level texture distribution pattern. At the local level, the texture of microcalcification clusters was aligned with the local peaks of the metabolomics signal to generate a local texture distribution pattern. Task-aware filtering was performed on cross-modal correlation features based on gradient contribution weights, retaining the features that contributed most to the thyroid nodule localization and malignancy prediction tasks. The filtered cross-modal correlation features were nonlinearly superimposed with the thyroid anatomical deformation field to generate an initial multi-omics mass spectrum. The target multi-omics mass spectrum was generated by iteratively optimizing the spatial coordinate error between the metabolomics signal hotspots and the microcalcification clusters.

[0104] By constructing a three-dimensional spatial weight matrix and multi-level spatial decomposition, efficient fusion of thyroid anatomical structure characteristics and multi-omics mass spectrometry data is achieved; through task-aware filtering and nonlinear superposition, the spatial alignment accuracy of cross-modal correlation features and anatomical structure deformation fields is optimized; the target multi-omics mass spectrometry map finally generated improves the comprehensiveness and accuracy of thyroid disease diagnosis and solves the problems of isolated multimodal data and insufficient spatial alignment accuracy.

[0105] In order to further improve the accuracy and efficiency of multimodal data fusion, this application dynamically adjusts the spatial position of mass spectrometry detection points within the coordinate mapping range of each level based on the hierarchical resolution of thyroid ultrasound images, where the coordinate mapping range of the global level covers the overall outline of the thyroid gland and the topological boundaries of adjacent organs, and the mapping range of the local level is limited to the local spatial neighborhood of the microcalcification cluster; at each level of resolution, according to the spatial distance between the adjusted mass spectrometry detection point and the ultrasound image voxel, the metabolomics signal intensity distribution in the multi-omics mass spectrometry data is spatially interpolated and compensated to solve the problem of insufficient spatial resolution of the mass spectrometry data; according to The consistency between the spatial distribution path of proteomic marker expression and the directional characteristics of the texture distribution pattern is used to directionally enhance the texture features and improve the pathological information expression ability of the texture features. The compensated metabolomics signal intensity distribution and the enhanced texture features are channel-spliced at hierarchical resolution to generate cross-modal correlation features. The global-level splicing channel is used to calculate the covariance matrix of the overall deformation trend of the thyroid gland and the global distribution of the metabolomics signal. The local-level splicing channel is used to calculate the spatial offset between the texture details of the microcalcification cluster and the peak position of the metabolic signal, thereby realizing efficient fusion and spatial alignment of multimodal data.

[0106] In some embodiments, step 202 of fusing the texture distribution pattern with the multi-omics mass spectrometry data layer by layer to generate cross-modal correlation features includes:

[0107] Step 301: Analyze the multi-omics mass spectrometry data to obtain mass spectrometry detection points. Based on the hierarchical resolution of the thyroid ultrasound image, dynamically adjust the mass spectrometry detection points within the coordinate mapping range of each level, where the global coordinate mapping range covers the overall contour of the thyroid gland and the topological boundaries of adjacent organs, and the local mapping range is limited to the local spatial neighborhood of the microcalcification cluster in the thyroid ultrasound image.

[0108] In step 301, mass spectrometry detection points refer to detection points in the multi-omics mass spectrometry data collected by mass spectrometry equipment, including three-dimensional spatial coordinates and biomolecular information (such as metabolomics signal intensity and proteomics marker expression). Hierarchical resolution refers to the multi-scale resolution of thyroid ultrasound images, including global resolution (low resolution), mesoscopic resolution (medium resolution), and local resolution (high resolution). The coordinate mapping range refers to the spatial distribution of mass spectrometry detection points within the thyroid region. The global resolution covers the overall thyroid contour and the topological boundaries of adjacent organs, while the local resolution is limited to the local spatial neighborhood of microcalcification clusters. Adjacent organs to the thyroid gland include the carotid arteries, located on both sides of the thyroid gland and providing blood supply; the trachea, located posterior to the thyroid gland and in close contact with it; the esophagus, located posterior to the thyroid gland and adjacent to part of the thyroid gland; the recurrent laryngeal nerve, which passes behind the thyroid gland and controls vocal cord movement; and the parathyroid glands, located dorsally of the thyroid gland and responsible for regulating calcium and phosphorus metabolism. The adjacent range generally refers to organs and tissues that are in direct contact with the thyroid gland or within 10 mm of the thyroid margin.

[0109] In an embodiment of the present application, multi-omics mass spectrometry data is first analyzed to obtain mass spectrometry detection points. Then, based on the hierarchical resolution of the thyroid ultrasound image, the spatial position of the mass spectrometry detection points is dynamically adjusted within the coordinate mapping range of each level. Specifically, at the global level, the coordinate mapping range of the mass spectrometry detection points covers the topological boundaries of the overall outline of the thyroid gland and adjacent organs (such as the carotid artery and trachea); at the local level, the coordinate mapping range of the mass spectrometry detection points is limited to the local spatial neighborhood of the microcalcification cluster. By dynamically adjusting the coordinate mapping range of the mass spectrometry detection points, the spatial alignment accuracy of the multimodal data is ensured.

[0110] Step 302: At each level of resolution, perform spatial interpolation compensation on the metabolomics signal intensity distribution in the multi-omics mass spectrometry data according to the adjusted spatial distance between the mass spectrometry detection point and the ultrasound image voxel to obtain a compensated metabolomics signal intensity distribution.

[0111] In step 302, spatial interpolation compensation involves interpolating the metabolomics signal intensity distribution based on the spatial distance between the mass spectrometry detection point and the ultrasound image voxel, addressing the issue of insufficient spatial resolution in the mass spectrometry data. An ultrasound image voxel is the smallest spatial unit in a thyroid ultrasound image, containing three-dimensional coordinates and texture intensity information.

[0112] In the present embodiment, at each level of resolution, spatial interpolation compensation is performed on the metabolomics signal intensity distribution in the multi-omics mass spectrometry data based on the adjusted spatial distance between the mass spectrometry detection point and the ultrasound image voxel. Specifically, the attenuation coefficient of the metabolomics signal is corrected by calculating the texture intensity gradient direction of the ultrasound image voxel within a preset neighborhood of the mass spectrometry detection point, and then a compensated metabolomics signal intensity distribution is generated.

[0113] Step 303: Based on the spatial consistency between the spatial distribution path of the proteomic marker expression amount in the multi-omics mass spectrometry data and the directional characteristics of the texture distribution pattern, the texture characteristics of the texture distribution pattern are directionally enhanced to obtain enhanced texture characteristics.

[0114] In step 303, the directional characteristics of the texture distribution pattern refer to the spatial extension direction of texture features (such as hypoechoic areas and microcalcification clusters) in the thyroid ultrasound image. Directional enhancement involves enhancing the texture features based on the consistency between the spatial distribution path of proteomic marker expression and the directional characteristics of the texture distribution pattern, thereby improving the texture features' ability to convey pathological information. Proteomic marker expression refers to the expression level of specific proteins in thyroid tissue and is often associated with disease states (such as malignant tumors).

[0115] In the present embodiment, the texture features are directional-enhanced based on the consistency between the spatial distribution path of proteomic marker expression and the directional characteristics of the texture distribution pattern. Specifically, the directional information of the texture features is enhanced by matching the location of protein marker expression mutations with the spatial distribution path of blood flow signals in ultrasound images, thereby generating enhanced texture features.

[0116] Step 304: Channel-joining the compensated metabolomics signal intensity distribution and the enhanced texture features according to the hierarchical resolution to generate cross-modal correlation features, wherein the global-level joint channel contains the covariance matrix of the overall thyroid deformation trend and the global distribution of the metabolomics signal, and the local-level joint channel contains the texture details of the microcalcification cluster and the spatial offset of the peak position of the metabolic signal.

[0117] In step 304, a specific numerical example of channel stitching is as follows: assuming data for the compensated metabolomics signal intensity distribution (global level): matrix size: 64 × 64 × 32 (x, y, z coordinates); example value: the signal intensity at coordinates (10, 20, 5) is 1200. Enhanced texture features (global level): matrix size: 64 × 64 × 32 (x, y, z coordinates); example value: the texture intensity at coordinates (10, 20, 5) is 0.8. Level resolution: global level: 64 × 64 × 32; local level: 128 × 128 × 64. Global level stitching: The compensated metabolomics signal intensity distribution and the enhanced texture features are stitched at the global level resolution. For example, at coordinates (10, 20, 5), the metabolomics signal intensity is 1200 and the texture intensity is 0.8; the cross-modal correlation feature value after splicing is (1200, 0.8).

[0118] In this example, the compensated metabolomics signal intensity distribution and enhanced texture features are spliced together at hierarchical resolution to generate cross-modal correlation features. Specifically, at the global level, the spliced channels are used to calculate the covariance matrix between the overall thyroid deformation trend and the global distribution of metabolomics signals; at the local level, the spliced channels are used to calculate the spatial offset between the texture details of microcalcification clusters and the peak positions of metabolic signals.

[0119] Here's a specific example:

[0120] Assuming the physical spatial extent of the thyroid region is 0–100 mm along the x-axis, 0–100 mm along the y-axis, and 0–50 mm along the z-axis, a 100×100×50 three-dimensional spatial weight matrix was constructed. At the global level, the coordinate mapping range of the mass spectrometry detection points was mapped to cover the overall contour of the thyroid gland and the topological boundaries of adjacent organs (such as the carotid artery and trachea). At the local level, the coordinate mapping range of the mass spectrometry detection points was restricted to the local spatial neighborhood of the microcalcification cluster. At each resolution level, the metabolomics signal intensity distribution was spatially interpolated and compensated based on the spatial distance between the mass spectrometry detection points and the ultrasound image voxels to generate a compensated metabolomics signal intensity distribution. Based on the consistency between the spatial distribution path of proteomic marker expression and the directional characteristics of the texture distribution pattern, the texture features were directionally enhanced to generate an enhanced texture feature. The compensated metabolomics signal intensity distribution and the enhanced texture feature were then channel-wise concatenated at each resolution level to generate a cross-modal correlation feature.

[0121] By dynamically adjusting the coordinate mapping range of mass spectrometry detection points, the spatial alignment accuracy of multimodal data is ensured; through spatial interpolation compensation and directional enhancement processing, the pathological information expression ability of metabolomics signals and texture features is improved; through channel splicing to generate cross-modal correlation features, efficient fusion and spatial alignment of multimodal data are achieved, thereby improving the comprehensiveness and accuracy of thyroid disease diagnosis.

[0122] To address the issue of insufficient accuracy in the fusion of thyroid anatomical deformation features with multimodal data, this application decomposes the thyroid anatomical deformation field into local deformation parameters and global deformation trends within the thyroid tissue. The local deformation parameters are generated by quantifying the relative displacement between the thyroid margin and the carotid artery contact surface, and the global deformation trend is generated by extracting the spatial offset direction of the overall thyroid contour. Based on the local deformation parameters, the filtered cross-modal correlation features are spatially transformed, and the three-dimensional coordinate mapping relationship of the mass spectrometry detection points is dynamically adjusted. Based on the global deformation trend, the cross-modal correlation features after spatial coordinate transformation are weighted between levels, and the feature contribution ratio of the global level to the local level is dynamically adjusted according to the overlap between the spatial density distribution of microcalcification clusters and the metabolomics signal hotspot area. The weighted cross-modal correlation features are nonlinearly superimposed with the deformation parameters of the thyroid anatomical deformation field to generate an initial multi-omics mass spectrum map. The spatial coordinate errors between the metabolomics signal hotspot areas and the microcalcification clusters are iteratively optimized to correct the spatial alignment accuracy of the initial map, thereby achieving efficient fusion and spatial alignment of multimodal data.

[0123] In some embodiments, the nonlinear superposition of the filtered cross-modal correlation features with the thyroid anatomical structure deformation field to generate an initial multi-omics mass spectrum in step 204 includes:

[0124] Step 401: Decomposing the deformation field of the thyroid anatomical structure into local deformation parameters and global deformation trends within the thyroid tissue.

[0125] In step 401, the local deformation parameter refers to the deformation of a local region within the thyroid tissue, generated by quantifying the relative displacement between the thyroid margin and the interface between the carotid artery. The global deformation trend refers to the spatial offset direction of the overall thyroid contour, generated by extracting the deformation characteristics of the overall thyroid contour.

[0126] In this embodiment, the deformation field of the thyroid anatomy is decomposed into local deformation parameters and global deformation trends. Specifically, the local deformation parameters are generated by quantifying the relative displacement between the thyroid margin and the carotid artery interface, reflecting the deformation characteristics of the local region; the global deformation trend is generated by extracting the spatial offset direction of the overall thyroid contour, reflecting the overall deformation trend of the thyroid gland.

[0127] Step 402: Based on the local deformation parameters, perform spatial coordinate transformation on the filtered cross-modal correlation features to obtain cross-modal correlation features after spatial coordinate transformation.

[0128] In step 402 , spatial coordinate conversion refers to coordinate mapping the filtered cross-modal correlation features according to local deformation parameters to adjust the three-dimensional coordinate position of the mass spectrum detection point.

[0129] In the embodiments of the present application, spatial coordinate transformation is performed on the filtered cross-modal correlation features based on local deformation parameters. Specifically, by matching the texture extension direction of the low-echo region in the thyroid ultrasound image with the metabolomics signal gradient direction of the cross-modal correlation feature tensor, the three-dimensional coordinate mapping relationship of the mass spectrometry detection point is dynamically adjusted to generate the cross-modal correlation features after spatial coordinate transformation.

[0130] Step 403: Based on the global deformation trend, weight distribution is performed between levels on the cross-modal correlation features after the spatial coordinate transformation to obtain cross-modal correlation features after weight distribution.

[0131] In step 403, the inter-level weight distribution is to adjust the feature contribution ratio between the global level and the local level according to the overlap between the spatial density distribution of the microcalcification cluster and the metabolomics signal hotspot area. The spatial coordinates of the microcalcification cluster are the basic data of the spatial density distribution of the microcalcification cluster.

[0132] In this embodiment, based on the global deformation trend, the cross-modal correlation features after spatial coordinate transformation are weighted across different levels. Specifically, based on the spatial density distribution of microcalcification clusters in thyroid tissue and the overlap between the metabolomics signal hotspots, the contribution ratio of the global and local levels is dynamically adjusted to generate the weighted cross-modal correlation features.

[0133] Step 404: Nonlinearly superimpose the weighted cross-modal correlation features with the deformation parameters of the thyroid anatomical structure deformation field, and iteratively optimize the spatial coordinates of the metabolomics signal hotspot area and the microcalcification cluster to generate an initial multi-omics mass spectrum.

[0134] In step 404, the superposition formula example is: Superposition result = cross-modal correlation feature × deformation parameter + deformation parameter^2. 1. Input data: Cross-modal correlation feature: Global level: thyroid overall contour deformation trend (eigenvalue: 0.8); Local level: microcalcification cluster texture details (eigenvalue: 0.6). Deformation parameters of the thyroid anatomical deformation field: Global deformation trend: 0.7; Local deformation amplitude: 0.5. Spatial coordinate error between the metabolomics signal hotspot and the microcalcification cluster: Initial error: 5 mm. 2. Nonlinear Superposition: Global superposition results: 0.8 × 0.7 + 0.72 = 0.56 + 0.49 = 1.05; local superposition results: 0.6 × 0.5 + 0.52 = 0.30 + 0.25 = 0.55; 0.6 × 0.5 + 0.52 = 0.30 + 0.25 = 0.55. 3. Iterative Optimization: First iteration: Spatial coordinate error: 5 mm; Based on the error gradient, the coordinate offset of the mass spectrometry detection point was adjusted to reduce the error to 3 mm. Second iteration: Spatial coordinate error: 3 mm; Further adjustment of the coordinate offset reduced the error to 1 mm. Third iteration: Spatial coordinate error: 1 mm; Iterations were terminated when the error met the preset tolerance (<2 mm). The initial multi-omics mass spectrum was generated. Global level: overlay result 1.05; local level: overlay result 0.55; spatial coordinate error: 1 mm (satisfies the preset tolerance range). The initial multi-omics mass spectrum includes the covariance matrix of the global thyroid contour deformation trend and the global distribution of metabolomics signals, the spatial offset of the microcalcification cluster texture details and the metabolomics signal peak position at the local level, and the spatial coordinate error between the metabolomics signal hotspots and microcalcification clusters. Deformation parameters include the deformation trend of the overall thyroid contour (global deformation) and the deformation amplitude of local regions (local deformation).

[0135] In this embodiment, the weighted cross-modal correlation features are nonlinearly superimposed with the deformation parameters of the thyroid anatomical deformation field to generate an initial multi-omics mass spectrum. Specifically, the initial spectrum is generated by weighted fusion of the cross-modal correlation features and the deformation parameters. The spatial alignment accuracy of the initial spectrum is then corrected by iteratively optimizing the spatial coordinate errors between the metabolomics signal hotspots and the microcalcification clusters.

[0136] Here's a specific example:

[0137] Assuming the physical spatial extent of the thyroid region is 0–100 mm along the x-axis, 0–100 mm along the y-axis, and 0–50 mm along the z-axis, a 100×100×50 three-dimensional spatial weight matrix was constructed. At the global level, the overall thyroid contour was aligned with the global distribution of the metabolomics signal to generate a global texture distribution pattern. At the local level, the texture of microcalcification clusters was aligned with the local peaks of the metabolomics signal to generate a local texture distribution pattern. Task-aware filtering was performed on cross-modal correlation features based on gradient contribution weights, retaining the features that contributed most to the thyroid nodule localization and malignancy prediction tasks. The filtered cross-modal correlation features were nonlinearly superimposed with the thyroid anatomical deformation field to generate an initial multi-omics mass spectrum. The target multi-omics mass spectrum was generated by iteratively optimizing the spatial coordinate error between the metabolomics signal hotspots and the microcalcification clusters.

[0138] By decomposing the deformation field of the thyroid anatomical structure into local deformation parameters and global deformation trends, a quantitative description of the deformation characteristics of thyroid tissue is achieved; through spatial coordinate transformation and inter-level weight distribution, the spatial alignment accuracy of cross-modal correlation features is optimized; through nonlinear superposition and iterative optimization, a high-precision initial multi-omics mass spectrometry map is generated, which improves the comprehensiveness and accuracy of thyroid disease diagnosis.

[0139] In order to further improve the accuracy and efficiency of thyroid disease diagnosis, this application decomposes the thyroid anatomical structure in the target multi-omics mass spectrometry map into global hierarchical anatomical deformation features and local hierarchical microcalcification cluster density features, wherein the global hierarchical anatomical deformation features are generated by quantifying the displacement gradient direction of the contact surface between the thyroid edge and the carotid artery, and the local hierarchical microcalcification cluster density features are obtained by statistically analyzing the three-dimensional coordinate distribution density of the microcalcification clusters in the preset spatial neighborhood; the candidate thyroid nodule area is extracted from the overlapping area between the metabolomics signal hotspot area and the local hierarchical microcalcification cluster density feature, and the metabolomics signal gradient direction and texture distribution pattern are combined to form a single-layer anatomical deformation feature. The consistency coefficient between the directions is used to generate the initial boundary outline; based on the expression of proteomic markers and combined with the global hierarchical anatomical deformation characteristics, a set of mutation positions related to the preset malignant risk is screened from the candidate thyroid nodule area to generate protein expression abnormality hotspots; according to the spatial angle between the metabolomics signal gradient direction and the corresponding direction of the texture distribution pattern, the confidence of the malignant risk direction of the candidate thyroid nodule area is calculated; the spatial distribution path of the protein expression abnormality hotspot is superimposed with the confidence of the malignant risk direction to generate the three-dimensional positioning coordinates of the thyroid nodule and the malignancy probability prediction value, thereby realizing high-precision diagnosis and risk assessment of thyroid diseases.

[0140] In some embodiments, the step 105 includes performing thyroid pathology feature analysis on the target multi-omics mass spectrum to obtain thyroid pathology analysis results, which include the three-dimensional positioning coordinates of the thyroid nodule and the predicted value of the malignancy probability, including:

[0141] Step 501: Decompose the thyroid anatomical structure in the target multi-omics mass spectrometry map into global hierarchical anatomical deformation features and local hierarchical microcalcification cluster density features. During the decomposition process, the global hierarchical anatomical deformation features are generated by quantifying the displacement gradient direction of the contact surface between the thyroid margin and the carotid artery, and the local hierarchical microcalcification cluster density features are obtained by statistically analyzing the three-dimensional coordinate distribution density of microcalcification clusters within a preset spatial neighborhood.

[0142] In step 501, the global anatomical deformation feature refers to the deformation trend of the overall thyroid contour. It is generated by quantifying the displacement gradient direction of the interface between the thyroid margin and the carotid artery and is used to describe the overall deformation characteristics of the thyroid gland. The local microcalcification cluster density feature refers to the density distribution of microcalcification clusters within the thyroid tissue. It is obtained by statistically analyzing the three-dimensional coordinate distribution density of microcalcification clusters within a preset spatial neighborhood and is used to describe local pathological characteristics.

[0143] In this embodiment, the thyroid anatomy in the target multi-omics mass spectrometry profile is decomposed into global anatomical deformation features and local microcalcification cluster density features. Specifically, the global anatomical deformation features are generated by quantifying the displacement gradient direction of the interface between the thyroid margin and the carotid artery, reflecting the overall deformation trend of the thyroid gland; the local microcalcification cluster density features are obtained by statistically analyzing the three-dimensional coordinate distribution density of microcalcification clusters within a preset spatial neighborhood, reflecting the local pathological characteristics.

[0144] Step 502: Extract a candidate thyroid nodule region from the overlapping area between the metabolomics signal hotspot region and the local hierarchical microcalcification cluster density feature, and generate an initial boundary contour based on the candidate thyroid nodule region and the consistency coefficient between the metabolomics signal gradient direction and the texture distribution pattern direction.

[0145] In step 502, the thyroid nodule candidate region refers to the potential nodule region extracted from the overlapping region between the metabolomics signal hotspot region and the local hierarchical microcalcification cluster density feature. The initial boundary contour refers to the preliminary boundary shape extracted from the thyroid nodule candidate region, which is used to describe the initial spatial extent of the nodule.

[0146] In the present embodiment, candidate thyroid nodule regions were extracted from the overlapping areas between the metabolomics signal hotspots and the local hierarchical microcalcification cluster density features. Then, the initial boundary outline was generated by combining the consistency coefficient between the metabolomics signal gradient direction and the texture distribution pattern direction. Specifically, the position and shape of the nodule boundary were determined by calculating the spatial angle between the metabolomics signal gradient direction and the texture extension direction.

[0147] Step 503: Based on the expression level of the proteomic marker and in combination with the global hierarchical anatomical deformation characteristics, a set of mutation positions associated with a preset malignancy risk is screened from the candidate thyroid nodule region, and based on the set of mutation positions, a protein expression abnormality hotspot is generated.

[0148] In step 503, the set of mutation locations refers to mutation locations of proteomic markers associated with malignancy risk, screened from candidate thyroid nodules. For example, when the correlation coefficient between protein marker expression and anatomical deformation magnitude is greater than a preset threshold (e.g., 0.7), the location is considered a mutation location.

[0149] In this example, based on the expression of proteomic markers and combined with global hierarchical anatomical deformation characteristics, a set of mutation locations associated with a predetermined malignancy risk was screened from candidate thyroid nodules. Spatial clustering analysis was then performed on this set of mutation locations to generate protein expression hotspots. Specifically, by calculating the spatial distance and expression similarity between mutation locations, mutation locations that met the criteria were clustered into a single aberration hotspot.

[0150] Step 504: Calculate the malignancy risk direction confidence of the thyroid nodule candidate region according to the spatial angle between the metabolomics signal gradient direction and the texture distribution pattern.

[0151] In step 504, the confidence in the malignant risk direction refers to a confidence index for evaluating the malignant risk direction of the candidate thyroid nodule region based on the degree of consistency between the metabolomics signal gradient direction and the texture distribution pattern direction. It is used to evaluate the malignant risk direction of the candidate thyroid nodule region and support the generation of a malignant probability prediction value.

[0152] In the present embodiment, the malignancy risk direction confidence of the thyroid nodule candidate region is calculated based on the spatial angle between the metabolomics signal gradient direction and the corresponding direction of the texture distribution pattern. Specifically, the malignancy risk direction confidence is generated by calculating the cosine value of the angle between the two.

[0153] Step 505: Superimpose the spatial distribution path of the abnormal protein expression hotspot and the malignancy risk direction confidence to generate the three-dimensional positioning coordinates of the thyroid nodule and the malignancy probability prediction value.

[0154] In this embodiment of the present application, the spatial distribution path of abnormal protein expression hotspots is superimposed with the confidence level of the malignancy risk direction to generate the three-dimensional location coordinates of the thyroid nodule and the malignancy probability prediction value. Specifically, the spatial distribution path of abnormal protein expression hotspots and the confidence level of the malignancy risk direction are weighted and fused to generate the superposition result; then, based on the superposition result, the three-dimensional location coordinates of the thyroid nodule and the malignancy probability prediction value are generated.

[0155] Here's a specific example:

[0156] Assuming the physical spatial extent of the thyroid region is 0–100 mm along the x-axis, 0–100 mm along the y-axis, and 0–50 mm along the z-axis, a 100×100×50 three-dimensional spatial weight matrix was constructed. At the global level, the overall thyroid contour was aligned with the global distribution of the metabolomics signal to generate a global texture distribution pattern. At the local level, the microcalcification cluster texture was aligned with the local peak of the metabolomics signal to generate a local texture distribution pattern. Thyroid nodule candidate regions were extracted from the overlapping regions between the metabolomics signal hotspots and the local microcalcification cluster density features. The initial boundary contours were generated by combining the consistency coefficient between the metabolomics signal gradient direction and the texture distribution pattern direction. Based on the expression levels of proteomic markers and the global anatomical deformation characteristics, a set of mutation locations associated with the pre-defined malignancy risk was screened from the thyroid nodule candidate regions to generate protein expression abnormality hotspots. The confidence level of the malignancy risk direction of the thyroid nodule candidate regions was calculated based on the spatial angle between the metabolomics signal gradient direction and the corresponding direction of the texture distribution pattern. The spatial distribution path of abnormal protein expression hotspots is superimposed with the confidence level of the malignant risk direction to generate the three-dimensional positioning coordinates of thyroid nodules and the predicted value of malignancy probability.

[0157] By decomposing the thyroid anatomical structure into global and local hierarchical features, a quantitative description of thyroid tissue deformation and pathological characteristics is achieved; by extracting candidate thyroid nodule regions and generating initial boundary contours, the accuracy of nodule positioning is optimized; by screening mutation position sets and generating abnormal protein expression hotspots, the accuracy of malignancy risk assessment is improved; by calculating the confidence level of the malignancy risk direction and superimposing the spatial distribution path, the three-dimensional positioning coordinates of the thyroid nodule and the malignancy probability prediction value are generated, thereby improving the comprehensiveness and accuracy of thyroid disease diagnosis.

[0158] In order to solve the accuracy problem of precise positioning of thyroid nodules and prediction of malignancy probability, this application first generates a spatial distribution path of protein expression abnormality hotspots by calculating the similarity between the spatial distance between mutation positions in the mutation position set and the expression amount of proteomic markers. This process uses the spatial clustering method to aggregate mutation positions with similar expression amounts and close spatial distances into abnormal hotspots; then, the protein expression abnormality hotspot intensity value of each point on the spatial distribution path is superimposed with the malignancy risk direction confidence to generate a superposition result, where the malignancy risk direction confidence is calculated based on the consistency between the metabolomics signal gradient direction and the texture distribution pattern direction; finally, based on the superposition result, the three-dimensional positioning coordinates of the thyroid nodule are determined by iteratively optimizing the coupling error between the spatial distribution path and the initial boundary contour, and at the same time, the spatial correlation between the spatial distribution path and the texture distribution pattern is verified to generate a malignancy probability prediction value of the thyroid nodule, thereby achieving precise positioning of thyroid nodules and malignancy risk assessment.

[0159] In some embodiments, the step 505 of superimposing the spatial distribution path of the abnormal protein expression hotspot and the malignancy risk direction confidence to generate the three-dimensional positioning coordinates of the thyroid nodule and the malignancy probability prediction value includes:

[0160] Step 601: Generate the spatial distribution path of the abnormal protein expression hotspot by calculating the similarity between the spatial distances between the mutation positions in the mutation position set and the expression levels of the proteomic markers.

[0161] In step 601, spatial distance refers to the Euclidean distance between mutation positions, which is used to describe the spatial distribution of mutation positions. The spatial distribution path of protein expression aberration hotspots refers to the spatial distribution trajectory of protein expression aberration hotspots generated through spatial clustering analysis, reflecting the spatial clustering pattern of protein marker expression mutation positions. A specific example is given: for each pair of mutation positions in a mutation position set, the spatial distance and expression similarity between them are calculated. For example, the spatial distance between mutation positions A and B is 7.35 mm, and the expression similarity is 0.92. Based on the spatial distance and expression similarity, spatial clustering analysis is performed on the mutation positions to generate protein expression aberration hotspots. For example, mutation positions with a spatial distance less than 10 mm and an expression similarity greater than 0.9 are clustered into one aberration hotspot: A, B, and C are clustered into one aberration hotspot; D and E are clustered into another aberration hotspot. By connecting multiple aberration hotspots, the spatial distribution path of protein expression aberration hotspots is generated. For example, the spatial distance and expression similarity between abnormal hotspots ABC and DE are calculated. If the preset conditions are met (such as the spatial distance is less than 20 mm and the expression similarity is greater than 0.8), the two are connected to generate a spatial distribution path.

[0162] In the examples of this application, the spatial distances between mutation positions in a set of mutation positions are first calculated, for example, using the Euclidean distance formula to calculate the distance between each pair of mutation positions. The similarity of proteomic marker expression between the mutation positions is then calculated, for example, by calculating a similarity coefficient by normalizing the expression difference. Based on the spatial distances and expression similarities, a spatial clustering algorithm is then used to cluster the mutation positions into hotspots of abnormal protein expression, generating a spatial distribution path. Ultimately, the spatial distribution path reflects the spatial clustering pattern of protein marker expression mutation positions in thyroid tissue.

[0163] Step 602: Superimpose the spatial distribution path and the malignant risk direction confidence to generate a superposition result.

[0164] In step 602, an example of the overlay process is provided: the spatial distribution path is represented as a series of spatial coordinate points and their corresponding protein expression abnormality hotspot intensities. For example, the spatial distribution path includes the following coordinate points: Point 1: (50, 60, 20), intensity 1200; Point 2: (55, 65, 22), intensity 1100; Point 3: (52, 63, 21), intensity 1150; Point 4: (70, 80, 25), intensity 900; Point 5: (75, 85, 27), intensity 950. The confidence level of the malignant risk direction is represented as the confidence value of each point on the spatial distribution path. For example, the confidence level of the malignant risk direction for each point on the spatial distribution path is 0.8. The protein expression abnormality hotspot intensities and the malignant risk direction confidence values for each point on the spatial distribution path are weighted and fused to generate the overlay result. For example, the value of each point in the overlay result is: Overlay result = protein expression abnormality hotspot intensity × malignant risk direction confidence.

[0165] In the present embodiment, the intensity value of each point on the spatial distribution path of the abnormal protein expression hotspot is first obtained, for example, by normalizing the protein marker expression level; then the confidence level of the malignant risk direction corresponding to each point is obtained, for example, by calculating the spatial angle between the metabolomics signal gradient direction and the texture distribution pattern direction; then, the intensity value of each point on the spatial distribution path and the confidence level of the malignant risk direction are weightedly superimposed to generate a superposition result. The superposition result reflects the spatial consistency of the abnormal protein expression hotspot and the malignant risk direction.

[0166] Step 603: Based on the superposition result, optimize the coupling error between the spatial distribution path and the initial boundary contour to determine the three-dimensional positioning coordinates of the thyroid nodule, verify the spatial correlation between the spatial distribution path and the texture distribution pattern, and generate a predicted value of the malignancy probability of the thyroid nodule.

[0167] In step 603, the coupling error, which refers to the spatial alignment error between the spatial distribution path and the initial boundary contour, is used to optimize the 3D positioning coordinates of the thyroid nodule. The spatial correlation, which refers to the degree of spatial consistency between the spatial distribution path of the abnormal protein expression hotspot and the texture distribution pattern, is used to generate a predicted value for the malignancy probability of the thyroid nodule.

[0168] In the embodiments of the present application, the three-dimensional positioning coordinates of the thyroid nodule are first determined based on the superposition results by iteratively optimizing the coupling error between the spatial distribution path and the initial boundary contour, for example, by minimizing the spatial alignment error using the gradient descent method. The spatial correlation between the spatial distribution path and the texture distribution pattern is then verified, for example, by calculating the spatial angle or overlap between the two, to generate a predicted value for the malignancy probability of the thyroid nodule. Ultimately, the three-dimensional positioning coordinates and the predicted value for the malignancy probability reflect the spatial location and malignancy risk of the thyroid nodule, respectively.

[0169] Here's a specific example:

[0170] Assume that a candidate thyroid nodule region contains five mutation sites. Mutation site A has three-dimensional coordinates of (50, 60, 20) and a proteomic marker expression level of 1200; mutation site B has three-dimensional coordinates of (55, 65, 22) and an expression level of 1100; mutation site C has three-dimensional coordinates of (52, 63, 21) and an expression level of 1150; mutation site D has three-dimensional coordinates of (70, 80, 25) and an expression level of 900; and mutation site E has three-dimensional coordinates of (75, 85, 27) and an expression level of 950. In step 601, the spatial distances between mutation sites are first calculated. For example, the distance between A and B is 7.35 mm. Expression level similarity is then calculated. For example, the similarity between A and B is 0.92. A spatial clustering algorithm is then used to cluster A, B, and C into one abnormal hotspot, and D and E into another abnormal hotspot, generating a spatial distribution path. In step 602, the intensity value of each point on the spatial distribution path is obtained (for example, the intensity value of point A is 1.0). The corresponding confidence level of the malignancy risk direction is also obtained for each point (for example, the confidence level of point A is 0.85). Finally, the intensity values and confidence levels are weighted and superimposed to generate a superposition result. In step 603, based on the superposition result, the coupling error between the spatial distribution path and the initial boundary contour is iteratively optimized to determine the three-dimensional positioning coordinates of the thyroid nodule as (53, 64, 21). The spatial correlation between the spatial distribution path and the texture distribution pattern is verified, and a predicted malignancy probability value of 0.78 is generated for the thyroid nodule.

[0171] By calculating the spatial distance between mutation locations and the similarity between proteomic marker expression levels, a spatial distribution path of abnormal protein expression hotspots is generated. This spatial distribution path is then superimposed with the confidence level of the malignancy risk direction to generate an overlay result. Based on this overlay result, the coupling error between the spatial distribution path and the initial boundary contour is optimized to determine the three-dimensional positioning coordinates of the thyroid nodule. The spatial correlation between the spatial distribution path and the texture distribution pattern is then verified to generate a predicted value for the malignancy probability of the thyroid nodule. This method achieves precise localization of thyroid nodules and prediction of malignancy probability, improving the accuracy and reliability of thyroid disease diagnosis.

[0172] In order to solve the gradient conflict problem between the thyroid nodule localization task and the malignancy prediction task in multi-task learning, this application first decomposes the back-propagation error of the thyroid anatomical structure deformation field into a back-propagation error component of the thyroid nodule localization task and a back-propagation error component of the thyroid nodule malignancy prediction task, wherein the error component of the localization task is generated by quantifying the displacement gradient direction of the interface between the thyroid margin and the carotid artery, while the error component of the malignancy prediction task is generated by quantifying the correlation between the position of the proteomic marker expression mutation and the anatomical deformation amplitude at the thyroid-carotid artery interface; then, based on the nonlinear relationship between the back-propagation error components of the two tasks, a nonlinear coupling function is constructed. The function is generated by fusing the contribution of the localization task error component to the overall contour deformation trend of the thyroid gland with the contribution of the malignancy prediction task error component to the spatial distribution path of the proteomic marker expression mutation position; finally, the nonlinear coupling function is applied to the back-propagation error of the thyroid anatomical structure deformation field, and by iteratively optimizing the coupling error of the two task error components, the gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task are calculated, thereby dynamically adjusting the optimization target weights of the two tasks during the iterative process of multi-task learning to achieve collaborative optimization between tasks.

[0173] In some embodiments, in step 103, during the multi-task learning iteration process, performing nonlinear coupling calculation on the back propagation error of the thyroid anatomical structure deformation field to obtain the respective gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task includes:

[0174] Step 701: Decomposing the back propagation error of the thyroid anatomical structure deformation field into a back propagation error component for the thyroid nodule localization task and a back propagation error component for the thyroid nodule malignancy prediction task.

[0175] In step 701, the backpropagation error refers to the error between the predicted and true values of the thyroid anatomical deformation field during the multi-task learning iteration process, calculated using the backpropagation algorithm. The backpropagation error component for the thyroid nodule localization task refers to the portion of the backpropagation error related to the thyroid nodule localization task and is generated by quantifying the displacement gradient direction at the interface between the thyroid margin and the carotid artery. The backpropagation error component for the thyroid nodule malignancy prediction task refers to the portion of the backpropagation error related to the thyroid nodule malignancy prediction task and is generated by quantifying the correlation between the location of proteomic marker expression mutations and the magnitude of anatomical deformation at the thyroid-carotid artery interface.

[0176] In this embodiment, the backpropagation error of the thyroid anatomical deformation field is first obtained, for example, by calculating the mean square error between the predicted value and the true value. The backpropagation error is then decomposed into a backpropagation error component for the thyroid nodule localization task and a backpropagation error component for the thyroid nodule malignancy prediction task. The error component for the localization task is generated by calculating the displacement gradient direction of the interface between the thyroid margin and the carotid artery, while the error component for the malignancy prediction task is generated by calculating the correlation between the location of proteomic marker expression mutations and the magnitude of anatomical deformation. Ultimately, the decomposed error components reflect the contributions of the localization task and the malignancy prediction task to the backpropagation error, respectively.

[0177] Step 702: Construct a nonlinear coupling function based on the nonlinear relationship between the back-propagation error component of the thyroid nodule localization task and the back-propagation error component of the thyroid nodule malignancy prediction task.

[0178] In step 702, the nonlinear coupling function refers to a function constructed based on the nonlinear relationship between the back-propagation error components of the two tasks, and is used to calculate the gradient contribution weights of the two tasks.

[0179] In an embodiment of the present application, the nonlinear relationship between the back-propagation error component of the thyroid nodule localization task and the back-propagation error component of the thyroid nodule malignancy prediction task is first analyzed, for example, by drawing a scatter plot of the error components or calculating the correlation coefficient; then a nonlinear coupling function is constructed based on the nonlinear relationship, for example, an exponential function or a logarithmic function is used to describe the interaction between the two error components; finally, a nonlinear coupling function is generated by fusing the contribution of the localization task error component to the overall contour deformation trend of the thyroid gland and the contribution of the malignancy prediction task error component to the spatial distribution path of the proteomic marker expression mutation position.

[0180] Step 703: Based on the nonlinear coupling function, the gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task are calculated.

[0181] In an embodiment of the present application, a nonlinear coupling function is first applied to the back-propagation error of the thyroid anatomical structure deformation field to calculate the gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task. For example, the weights are determined by iteratively optimizing the coupling error of the error components of the two tasks. Then, based on the gradient contribution weights, the optimization target weights of the two tasks are dynamically adjusted during the multi-task learning iteration process. For example, the optimization target weights are generated by matching the contribution ratios of the error components of the two tasks. Ultimately, the gradient contribution weights achieve collaborative optimization between the two tasks.

[0182] Here's a specific example:

[0183] Assume that the backpropagation error of the thyroid anatomical deformation field is 0.5, of which the backpropagation error component for the thyroid nodule localization task is 0.3, and the backpropagation error component for the thyroid nodule malignancy prediction task is 0.2. In step 701, the backpropagation error is first decomposed into the error component for the localization task (0.3) and the error component for the malignancy prediction task (0.2). The error component for the localization task is generated by calculating the displacement gradient direction at the interface between the thyroid margin and the carotid artery, while the error component for the malignancy prediction task is generated by calculating the correlation between the location of proteomic marker expression mutations and the magnitude of anatomical deformation. In step 702, a nonlinear coupling function is constructed based on the nonlinear relationship between the two error components. For example, an exponential function is used to describe the interaction between the two error components. In step 703, the nonlinear coupling function is applied to the backpropagation error. By iteratively optimizing the coupling error of the error components of the two tasks, the gradient contribution weights for the localization task are calculated to be 0.6, and the gradient contribution weight for the malignancy prediction task is calculated to be 0.4. The optimization target weights for the two tasks are dynamically adjusted based on the weights.

[0184] By decomposing the back-propagation error of the thyroid anatomical structure deformation field into the back-propagation error components of the thyroid nodule localization task and the malignancy prediction task, a nonlinear coupling function is constructed based on the nonlinear relationship between the two error components, and the gradient contribution weights of the two tasks are calculated. This achieves collaborative optimization between the thyroid nodule localization task and the malignancy prediction task in multi-task learning, thereby improving the accuracy and reliability of thyroid disease diagnosis.

[0185] Figure 2 The present invention provides a schematic diagram of a multi-omics mass spectrometry analysis system for thyroid disease. Figure 2 As shown, the system includes:

[0186] an acquisition module 21 for acquiring a thyroid ultrasound image and multi-omics mass spectrometry data, and generating a cross-modal association relationship based on the anatomical structural features of the thyroid ultrasound image and the spatial distribution features of biomolecules in the multi-omics mass spectrometry data;

[0187] A construction module 22 is configured to generate a thyroid anatomical structure deformation field based on the anatomical structure features, and to construct a multi-task optimization objective according to preset thyroid disease diagnosis requirements, wherein the multi-task optimization objective includes a thyroid nodule localization task and a thyroid nodule malignancy prediction task;

[0188] A calculation module 23 is configured to perform nonlinear coupling calculation on the back propagation error of the thyroid anatomical structure deformation field during the multi-task learning iteration process to obtain the gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task;

[0189] a generating module 24 for generating a cross-modal association feature according to the cross-modal association relationship, and generating a target multi-omics mass spectrum based on the thyroid anatomical structure deformation field and the cross-modal association feature in combination with the gradient contribution weight, wherein the error between the metabolomics signal hotspot region in the target multi-omics mass spectrum and the spatial coordinates of the microcalcification cluster in the thyroid ultrasound image is within a preset anatomical structure tolerance range;

[0190] The analysis module 25 is used to analyze the thyroid pathological characteristics of the target multi-omics mass spectrum to obtain thyroid pathological analysis results, which include the three-dimensional positioning coordinates of the thyroid nodules and the predicted value of the malignancy probability.

[0191] Figure 2 The multi-omics mass spectrometry analysis system for thyroid disease can be executed Figure 1 The implementation principles and technical effects of the multi-omics mass spectrometry analysis method for thyroid disease described in the illustrated embodiment will not be elaborated on here. The specific manner in which each module and unit performs operations in the multi-omics mass spectrometry analysis system for thyroid disease described in the above embodiment has been described in detail in the embodiments of the method and will not be elaborated on here.

[0192] In one possible design, Figure 2 The multi-omics mass spectrometry analysis system for thyroid disease of the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0193] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0194] The processing component 32 is as follows Figure 1 The embodiment provides a method for constructing and analyzing a multi-omics mass spectrometry map of thyroid disease.

[0195] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0196] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0197] Of course, a computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0198] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0199] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0200] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0201] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The embodiment shown is a method for constructing and analyzing a multi-omics mass spectrometry map of thyroid disease.

[0202] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0203] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0204] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0205] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for constructing and analyzing multi-omics mass spectrometry profiles of thyroid diseases, characterized in that: include: Acquiring a thyroid ultrasound image and multi-omics mass spectrometry data, and generating a cross-modal association relationship based on anatomical structural features of the thyroid ultrasound image and biomolecule spatial distribution features of the multi-omics mass spectrometry data, wherein the anatomical structure includes a thyroid nodule; generating a thyroid anatomical structure deformation field based on the anatomical structure features, and constructing a multi-task optimization target according to preset thyroid disease diagnosis requirements, wherein the multi-task optimization target includes a thyroid nodule localization task and a thyroid nodule malignancy prediction task; During the multi-task learning iteration process, a nonlinear coupling calculation is performed on the back propagation error of the thyroid anatomical structure deformation field to obtain the respective gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task; generating a cross-modal association feature according to the cross-modal association relationship, and generating a target multi-omics mass spectrum based on the thyroid anatomical structure deformation field and the cross-modal association feature in combination with the gradient contribution weight, wherein the error between the metabolomics signal hotspot area in the target multi-omics mass spectrum and the spatial coordinates of the microcalcification cluster in the thyroid ultrasound image is within a preset anatomical structure tolerance range; The target multi-omics mass spectrometry map is analyzed for thyroid pathological characteristics to obtain thyroid pathological analysis results, which include the three-dimensional positioning coordinates of the thyroid nodule and the predicted value of the malignancy probability.

2. The method according to claim 1, characterized in that Generating a cross-modal correlation feature according to the cross-modal correlation relationship, and generating a target multi-omics mass spectrum based on the thyroid anatomical structure deformation field and the cross-modal correlation feature in combination with the gradient contribution weight, including: Based on the cross-modal association relationship, the anatomical structure features and the biomolecule spatial distribution features are combined to construct a three-dimensional spatial weight matrix, and the anatomical structure features are subjected to multi-level spatial decomposition according to the three-dimensional spatial weight matrix to obtain texture distribution patterns at different anatomical levels in the thyroid tissue; fusing the texture distribution pattern with the multi-omics mass spectrometry data layer by layer to generate cross-modal correlation features; performing task-aware filtering on the cross-modal correlation feature according to the gradient contribution weight to obtain a filtered cross-modal correlation feature; Nonlinearly superimposing the filtered cross-modal correlation features with the thyroid anatomical structure deformation field to generate an initial multi-omics mass spectrum; The initial multi-omics mass spectrum is optimized to generate a target multi-omics mass spectrum.

3. The method according to claim 2, characterized in that The step of fusing the texture distribution pattern with the multi-omics mass spectrometry data layer by layer to generate cross-modal correlation features includes: Analyzing the multi-omics mass spectrometry data to obtain mass spectrometry detection points, and dynamically adjusting the mass spectrometry detection points within a coordinate mapping range of each level based on the hierarchical resolution of the thyroid ultrasound image, wherein the global coordinate mapping range covers the overall contour of the thyroid gland and the topological boundaries of adjacent organs, and the local level mapping range is limited to the local spatial neighborhood of the microcalcification cluster in the thyroid ultrasound image; At each level of resolution, performing spatial interpolation compensation processing on the metabolomics signal intensity distribution in the multi-omics mass spectrometry data according to the adjusted spatial distance between the mass spectrometry detection point and the ultrasound image voxel to obtain a compensated metabolomics signal intensity distribution, wherein the ultrasound image voxel is the smallest spatial unit of the thyroid ultrasound image; performing directionality enhancement processing on the texture features of the texture distribution pattern according to the spatial consistency between the spatial distribution path of the proteomic marker expression amount in the multi-omics mass spectrometry data and the directional features of the texture distribution pattern to obtain enhanced texture features; The compensated metabolomics signal intensity distribution and the enhanced texture features are channel-spliced at the hierarchical resolution to generate cross-modal correlation features, wherein the global-level splicing channel is used to calculate the covariance matrix of the overall thyroid deformation trend and the global distribution of the metabolomics signal, and the local-level splicing channel is used to calculate the texture details of the microcalcification cluster and the spatial offset of the metabolic signal peak position.

4. The method according to claim 2, characterized in that The nonlinear superposition of the filtered cross-modal correlation features with the thyroid anatomical structure deformation field to generate an initial multi-omics mass spectrum includes: Decomposing the deformation field of the thyroid anatomical structure into local deformation parameters and global deformation trends within the thyroid tissue; Based on the local deformation parameters, performing spatial coordinate transformation on the filtered cross-modal correlation features to obtain cross-modal correlation features after spatial coordinate transformation; Based on the global deformation trend, performing inter-level weight distribution on the cross-modal correlation features after the spatial coordinate transformation to obtain cross-modal correlation features after weight distribution; The weighted cross-modal correlation features are nonlinearly superimposed with the deformation parameters of the thyroid anatomical structure deformation field to generate an initial multi-omics mass spectrum map, wherein the initial multi-omics mass spectrum map is corrected by iteratively optimizing the spatial coordinates of the metabolomics signal hotspot area and the microcalcification cluster.

5. The method according to claim 1, wherein The target multi-omics mass spectrometry profile is subjected to thyroid pathology feature analysis to obtain thyroid pathology analysis results, which include the three-dimensional positioning coordinates of the thyroid nodule and the predicted value of the malignancy probability, including: Decomposing the thyroid anatomical structure in the target multi-omics mass spectrometry map into global hierarchical anatomical deformation features and local hierarchical microcalcification cluster density features. During the decomposition process, the global hierarchical anatomical deformation features are generated by quantifying the displacement gradient direction of the interface between the thyroid margin and the carotid artery, and the local hierarchical microcalcification cluster density features are obtained by statistically analyzing the three-dimensional coordinate distribution density of microcalcification clusters within a preset spatial neighborhood. Extracting a candidate thyroid nodule region from the overlapping region between the metabolomics signal hotspot region and the local hierarchical microcalcification cluster density feature, and generating an initial boundary contour based on the candidate thyroid nodule region and the consistency coefficient between the metabolomics signal gradient direction and the corresponding direction of the texture distribution pattern; Based on the expression levels of proteomic markers and in combination with the global hierarchical anatomical deformation characteristics, a set of mutation positions associated with a preset malignancy risk is screened from the candidate thyroid nodule region, and based on the set of mutation positions, a hotspot of abnormal protein expression is generated; Calculating the malignancy risk direction confidence of the thyroid nodule candidate region according to the spatial angle between the metabolomics signal gradient direction and the corresponding direction of the texture distribution pattern; The spatial distribution path of the abnormal protein expression hotspot is superimposed with the confidence of the malignant risk direction to generate the three-dimensional positioning coordinates of the thyroid nodule and the malignancy probability prediction value.

6. The method according to claim 5, characterized in that The step of superimposing the spatial distribution path of the abnormal protein expression hotspot and the malignancy risk direction confidence to generate the three-dimensional positioning coordinates of the thyroid nodule and the malignancy probability prediction value includes: Generate a spatial distribution path of the abnormal protein expression hotspot by calculating the similarity between the spatial distances between the mutation positions in the mutation position set and the expression amount of the proteomic marker; Superimposing the protein expression abnormality hotspot intensity value of each point on the spatial distribution path with the malignancy risk direction confidence to generate a superposition result; Based on the superposition result, the coupling error between the spatial distribution path and the initial boundary contour is optimized to determine the three-dimensional positioning coordinates of the thyroid nodule, and the spatial correlation between the spatial distribution path and the texture distribution pattern is verified to generate a predicted value of the malignancy probability of the thyroid nodule.

7. The method according to claim 1, characterized in that In the multi-task learning iteration process, nonlinear coupling calculation is performed on the back propagation error of the thyroid anatomical structure deformation field to obtain the gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task, including: Decomposing the back propagation error of the thyroid anatomical structure deformation field into a back propagation error component for the thyroid nodule localization task and a back propagation error component for the thyroid nodule malignancy prediction task; constructing a nonlinear coupling function based on a nonlinear relationship between a back-propagation error component of the thyroid nodule localization task and a back-propagation error component of the thyroid nodule malignancy prediction task; Based on the nonlinear coupling function, the gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task are calculated.

8. A multi-omics mass spectrometry analysis system for thyroid disease, characterized by: include: an acquisition module, configured to acquire thyroid ultrasound images and multi-omics mass spectrometry data, and generate a cross-modal association relationship based on the anatomical structural features of the thyroid ultrasound images and the spatial distribution features of biomolecules in the multi-omics mass spectrometry data; A construction module is used to generate a thyroid anatomical structure deformation field based on the anatomical structure features, and to construct a multi-task optimization target according to preset thyroid disease diagnosis requirements, wherein the multi-task optimization target includes a thyroid nodule localization task and a thyroid nodule malignancy prediction task; a calculation module, configured to perform nonlinear coupling calculation on the back propagation error of the thyroid anatomical structure deformation field during the multi-task learning iteration process, to obtain the gradient contribution weights of the thyroid nodule localization task and the thyroid nodule malignancy prediction task; a generation module, configured to generate a cross-modal association feature according to the cross-modal association relationship, and generate a target multi-omics mass spectrum based on the thyroid anatomical structure deformation field and the cross-modal association feature, in combination with the gradient contribution weight, wherein the error between the metabolomics signal hotspot area in the target multi-omics mass spectrum and the spatial coordinates of the microcalcification cluster in the thyroid ultrasound image is within a preset anatomical structure tolerance range; An analysis module is used to analyze the thyroid pathological characteristics of the target multi-omics mass spectrometry map to obtain thyroid pathological analysis results, which include the three-dimensional positioning coordinates of the thyroid nodules and the predicted value of the malignancy probability.

9. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a multi-omics mass spectrometry analysis method for thyroid disease as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, the method for constructing and analyzing a multi-omics mass spectrometry map of a thyroid disease according to any one of claims 1 to 7 is implemented.

Images